Fix append! for matrix bounds in batch models

Split append! into AbstractVector and AbstractMatrix methods. Matrix
bounds (e.g., lvar = zeros(10,4) for a 4-instance batch) are now
cat'd directly along dim 1 instead of being vec'd to a flat vector
and then incorrectly expanded, which produced (40,4) instead of (10,4).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: sshin23

src/nlp.jl

@@ -686,11 +686,15 @@ function append!(backend, a, b::Number, lb)
     return cat(a, new_part; dims = 1)
 end
 
-function append!(backend, a, b::AbstractArray, lb)
+function append!(backend, a, b::AbstractVector, lb)
     lb == 0 && return a
     arr = convert_array(b, backend)
-    col = vec(arr)
-    return cat(a, _expand_to_shape(col, _trailing_dims(a)); dims = 1)
+    return cat(a, _expand_to_shape(arr, _trailing_dims(a)); dims = 1)
+end
+function append!(backend, a, b::AbstractMatrix, lb)
+    lb == 0 && return a
+    arr = convert_array(b, backend)
+    return cat(a, arr; dims = 1)
 end
 
 function append!(backend, a, b::Base.Generator, lb)